Biology-Guided Prototype Booster: Enhancing Latent Representations of Foundation Models for Gene Expression Prediction

ICLR 2026 Conference Submission17957 Authors

Keywords: Deep Learning, Representation Learning, Biomedical Image

Abstract: Spatial transcriptomics (ST) is a cutting-edge technology that enables the measurement of gene expression while preserving spatial context and generating detailed tissue images. However, ST technology remains time-consuming and costly. The ability to predict ST gene markers of cancer from histology-grade H&E-stained tissue images is opening new horizons for precision and personalised pathology. Despite the success of foundation models in generating general-purpose embeddings of H&E-images, these representations are not optimized for gene expression prediction and lack task-specific adaptability. To address this limitation, we introduce Biology-Guided Prototype Booster (BP-Booster), leveraging biological prior knowledge to guide the construction and training of learnable prototypes for embedding reconstruction, thereby improving gene expression prediction. We demonstrate superior performance of BP-Booster across datasets, various cancer tissue types and different ST platforms. We also show that BP-Booster can flexibly integrate various foundation models to enhance their task-specific representations, enhancing explainability and applicability in clinically relevant tasks like predicting cancer biomarkers. Code will be released upon acceptance.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 17957